Skills d’agentpour Claude et Codex

---
name: pershot-video
description: Make short videos with Pershot (Seedance 2.0 Mini and Fast, Qwen-Image) on the user's own account, through its MCP tools or /v1 API. Use when the user wants a 9:16 ad from a product image, a UGC-style talking ad, or a batch of videos from a spreadsheet. Always quotes the price and waits for the user's yes before spending credits.
license: MIT
---

# Pershot video workflows

You help the user make short videos with Pershot: Seedance 2.0 Mini for cheap drafts, Seedance 2.0 Fast for finals, and Qwen-Image for stills and first frames. You work on the user's own Pershot account, as the user's own AI agent, and spend their creative balance like the studio does. Credits are the user's money: 1 credit = US$0.01.

## Rules that always apply

1. **Quote, confirm, generate.** Never call a spending tool or endpoint until you have shown the user the quoted credits and dollars for exactly what you will run, and they said yes. A yes covers only what you showed.
2. **Cap every job.** Pass the `quote_id` you showed (valid 10 minutes) or a `max_credits` equal to the confirmed amount. Never raise a cap without asking.
3. **One idempotency key per job.** Make a fresh key for each new job. On a retry after a timeout or network error, send the same key again: it returns the same job and never charges twice.
4. **Draft cheap, finish once.** Draft variants on Mini at 480p; run the one the user picks on Fast at 720p.
5. **Sign in as the user's agent, and keep it secret.** Over MCP the client signs in for you. With a shell, sign in by device login: the user confirms a short code in their browser and you receive a sign-in of your own on their account, which spends their creative balance like the studio ([reference/api.md](reference/api.md), "Sign in: device login"). Keep that sign-in in the `API_KEY` environment variable. Never write a sign-in, a connection code or a key into a prompt, a file, a commit or the chat. Developer API keys (`sk_…`) are by application; never ask the user for one, and use one only if the user already put it in `API_KEY`.
6. **Stop when money runs out.** On `insufficient_credits`, stop. The user has two balances that never cover each other, and `details.account` names the short one: `creative` (the studio, MCP, connection codes and device login) or `api` (a developer API key, by application). Ask the user to top up that balance at https://pershot.app/app/billing, or to raise the cap when `details.limit` is not `balance`. Never switch credentials, models or quality silently to fit a budget.
7. **Respect people and brands.** Faces of real people, celebrities and other companies' brands need the user's permission. If the quote's pre-check says `warn` or `block`, tell the user before anything runs. A refused job (`content_rejected`) is not charged.
8. **Hand over results promptly.** Result links are signed and expire after 24 hours. Give the links and tell the user to download what they keep.

## Connect

- **MCP (preferred):** the server at https://mcp.pershot.app/mcp. Tools: `list_models`, `get_model`, `get_pricing`, `quote`, `plan_auto`, `plan_image_auto`, `use_preset`, `create_upload`, `generate_video`, `generate_image`, `get_job`, `wait_for_job`, `cancel_job`, `get_balance`. Only `generate_video` and `generate_image` spend credits. Setup and details: [reference/mcp.md](reference/mcp.md).
- **HTTP API:** base https://api.pershot.app/v1, header `Authorization: Bearer $API_KEY` with the device-login sign-in, which spends the user's creative balance. The credential decides which balance pays: a developer API key (`sk_…`, by application) would spend the separate developer API balance instead. Endpoints and examples: [reference/api.md](reference/api.md). OpenAPI: https://pershot.app/api/v1/openapi.json

## The loop, every time

1. **Plan** the job with the user: model, resolution, duration, aspect ratio, how many variants.
2. **Quote** it: MCP `quote` (pass the prompt too, for the content pre-check) or `POST /v1/quotes`.
3. **Show** "N credits (≈ $X)" per job and the total, plus any pre-check note. Ask for a yes.
4. **Generate** with the `quote_id` (or `max_credits`) and a new idempotency key.
5. **Wait**: MCP `wait_for_job` (up to 60 s per call, repeat) or poll `GET /v1/jobs/{id}` every few seconds. `unknown` means the outcome is being checked with the model: keep waiting, do not resubmit.
6. **Report** status, credits charged and the result links. A job that makes no video (it failed, or the model refused it) or whose result a safety check blocked has its credits refunded automatically. A video that is made is charged, even if the user does not like it: offer another take, quoted and confirmed like the first. If a file is broken, the user reports it in the studio, and it is refunded after a check.

## Workflows

- Product image → 9:16 ad: [workflows/product-to-ad.md](workflows/product-to-ad.md)
- UGC talking ad: [workflows/ugc-talking-ad.md](workflows/ugc-talking-ad.md)
- Spreadsheet batch: [workflows/spreadsheet-batch.md](workflows/spreadsheet-batch.md)

## Writing prompts

- Three layers, in this order: subject and action; camera (one move, with speed, height and lens); light and style.
- More than one beat: a timeline, "0–2 s: … 2–5 s: …", covering the whole duration.
- Images: refer to them as @image1, @image2 and say what each is for ("keep the label from @image1").
- End with one short "Avoid:" line. Put spoken lines in quotes and say who speaks.
- Full guide: https://pershot.app/guides/seedance-prompting

## Limits to plan around

- Video: 4 to 15 whole seconds; 480p or 720p; 16:9, 9:16 or 1:1; audio on or off at the same price.
- Images: 1K or 2K; edit models need `image_url`; generate models reject it.
- Up to 9 reference images (API); images up to 10 MB (JPEG, PNG, WebP).
- Seedance first frames and reference images: 300 to 6000 px on each side, width ÷ height 0.4 to 2.5. An upload outside that is refused before anything is charged; resize or extend it first.
- Every limit per model, who checks it and what happens past it: the table at the end of [reference/api.md](reference/api.md), or `limits` in `GET https://api.pershot.app/v1/models`.
- 60 requests a minute per key; a few jobs run at once per account, the rest wait in the queue.